Background

Aesthetics plays an integral role in promoting personal well-being. While individuals may not be consciously aware of their choices, they intrinsically prefer a setting where they can function efficiently. Previous research showed that people have a preference for natural over artificial environments (Velarde, Fry, & Tveit, 2007; Berman, Jonides, & Kaplan, 2008). This aesthetic preference has been demonstrated to be strongly associated with nature’s potential restorative effects in the research (Purcell, Peron, & Berto, 2001; Hartig & Staats, 2006; Han, 2010). For example, previous research has shown salubrious effects after engaging with natural environment, such as improving memory, attention and mood (Berman et al., 2008; Berman et al., 2012). In modern times, however, increasing man-made architect and construction have alienated individuals from the natural environment. Therefore, it is important to gain a better understanding of people’s aesthetic preference of the environment in order to maintain a crucial engagement with nature and get potential benefits from it.

As driving has become a daily part of our everyday life, the field has yet, to date, studied people’s aesthetic preferences with regards to roadside environment. Past studies on scenic beauty suggested that people prefer the properties of nature. They reported that people prefer semantic features of nature, such as trees, water, and flowers (Nelson, 1997; Clay & Daniel, 2000; Brush, Chenoweth, & Barman, 2000), and that removal of built-up features like billboards could increase landscape appreciation (Antonson et al., 2009; Garré, Meeus, & Gulinck, 2009). A recent study (Kardan & Berman et al., 2015) quantified natural scenes by using low-level visual features (i.e., edge, hue, saturation, brightness, standard deviation of hue, standard deviation of saturation, etc.). In their study, participants were asked to rate the naturalness of the shown images and their likeness towards them. Results showed that low-level visual features significantly predicted people’s preference towards the images as well as the naturalness ratings of the images.

Current study

In the current study, we aim to investigate people’s aesthetic preference by using constructed highway environments with constantly changing surroundings. We generated simulation videos with the input of Geographical Information System (GIS) data, images and 3D models. The videos contain controlled environmental characteristics of a landscape, which are not only convenient for examining what specific design elements that affect scenic preference, but also simultaneously control for potential effects caused by demographic and social factors. Furthermore, we seek to provide a more nuanced understanding of preference by utilizing a slider bar to capture the continuous change of preference ratings throughout the whole experiment.

Hypotheses

We have several hypotheses. We hypothesize that both low-level visual features and semantic features of nature will be likely to predict preferences. Specifically, semantic features (e.g., trees and lake) that are related to naturalness will positively predict preference whereas artificial features (e.g., billboard) will negatively predict preference. With this study, we may be able to have a better understanding of people’s aesthetic preferences in terms of both low-level visual features and semantic features. This information could potentially be helpful for urban design and landscape planning in constructing environment that optimizes subjective well-being.

Participants

37 participants were recruited from the University of British Columbia, including 14 females and 23 males ranging from 20-59 years old. They came from a variety of disciplines and had various amount of education, ranging from technical diplomas to post-graduate level degrees.

Materials

We developed six custom GIS-based 4D real-time highway driving simulations where we controlled the type and location of environmental features (i.e., trees, lake, billboard, house and traffic signs), and topography (flat, rolling hills and mountainous forest). Each video was shown in the screen of roughly three meters high and four meters wide in order to represent landscape elements in a reasonably realistic scale. A slider bar was used to collect the real-time ratings throughout the video. As the movie progressed, participants were asked to lower or raise the value of the slider bar to indicate their preference. The ratings were collected on a scale from 0-10, where 10 corresponds to very high preference and 0 corresponds to very low preference. The ratings were taken every one-half second with a sensitivity of two decimal places.

Procedure

In the study, participants were asked to rate a series of highway landscapes represented by six movies (Flat 1, Flat 2, Hill 1, Hill 2, Mount 1, and Mount 2) using a slider bar. By adjusting the slider up or down, participants indicated their likeness towards the landscape more or less correspondingly. Before each video began, they were presented with nine photographs chosen from the upcoming video and they were asked to rate them using the slider bar. By doing this, we intended to familiarize the participants with the general landscape and the apparatus they were going use to rate the video. While participants were watching the video, they were asked to rate their preference towards the landscape in real time by adjusting the slider bar. After each video, they were also asked to report their overall impression of the landscape that they just traveled. Before a new video began, participants were asked to move the slider bar to the middle. Two videos of the same kind of topography would not be presented following one another. Six videos were presented in the same order to eliminate the potential order effects. Time to complete the entire study was approximately 45 minutes.

Overall plot

Coupling with snapshots from the videos

Results

The first figure shows the changes of preference ratings with the influences of different semantic features. Across six videos, videos of rolling hills and mountainous forests tended to have higher preferences than videos of flat terrains. The second figure links the first plot with exmaple scenes at some significant points.

As shown in the above figures, across six videos, billboard negatively impacted the preference ratings, whereas lake positively impacted the ratings. Trees sometimes contributed positively to the preference, and sometimes negatively impacted the preference, depending on scene types as well other factors, for example, slope or curves.

Visualization of the regression model – flat terrains

Results 1

In order to investigate how preference ratings were influenced under different features, we ran regression analyses for each scene type in a different model. Not all the semantic features were presented in videos of each scene type.

Flat terrains:
We ran a regression analysis for videos of flat terrains in which the preference ratings were regressed on all the low level visual features and the four semantic features (trees, traffic sign, house, and billboard). Low-level visual features and semantic features accounted for a significant proportion of variance in preference ratings (adjusted R square = .44, F(12,467) = 32.68, p < .001). In this regression model, some low-level visual features (hue, brightness, SD hue, SD saturation, and SD brightness) and semantic features (billboard and trees) significantly predicted people’s preference. Specifically, higher hue and variations in brightness, lower brightness and variations in hue and saturation might increase people’s preferences of the scene. Billboard and trees both negatively impacted the preference ratings.

Visualization of the regression model – rolling hills

Results 2

Rolling Hills:
We also ran a regression analysis for videos of rolling hills in which the preference ratings were regressed on all the low level visual features and the three semantic features (trees, traffic sign, and lake). Low-level visual features and semantic features accounted for a significant proportion of variance in preference ratings (adjusted R square = .32, F(11,468) = 21.93, p < .001). In this regression model, some low-level visual features (saturation, brightness, SD saturation, SD brightness, and entropy) and semantic features (trees and lake) significantly predicted people’s preference. Specifically, higher saturation and entropy, and variations in brightness, lower brightness and variations in saturation might increase people’s preferences of the scene. Lake positively impacted the preference ratings whereas trees still negatively impacted the preference ratings.

Visualization of the regression model – mountainous forests

Results 3

Mountainous forests:
We also ran a regression analysis for videos of mountainous forests in which the preference ratings were regressed on all the low level visual features and the two semantic features (trees and lake). Low-level visual features and semantic features accounted for a significant proportion of variance in preference ratings (adjusted R square = .68, F(10,457) = 101.5, p < .001). In this regression model, some low-level visual features (hue, brightness, and SD hue) and semantic features (trees and lake) significantly predicted people’s preference. Specifically, more variations in hue, lower hue and brightness might increase people’s preferences of the scene. Lake and trees both positively impacted the preference ratings.

Comparing between different models

Results 4

Although features were preferred different in different videos, we could still discover some consistencies between videos: more saturation and variations in brightness positively contributed to the preference ratings; but more brightness and variations in saturation were negatively contributed to the preference ratings. In terms of semantic features, billboard negatively impacted the preference, whereas lake positively impacted the preference.

Correlation map

Conclusions and discussion

We showed in our study that across six constructed environments low-level visual features and semantic features could significantly predict aesthetic preference. Our results are in general consistent with previous literature (Kardan & Berman et al., 2015) that the regression model containing low-level visual features would significantly predict preference. For each one of the regression models, low-level visual features were all shown to account for some of the variance in preference ratings. However, the predictive directions of those low-level visual features varied in each video. Our results did not suggest consistent predictive directions of those low-level visual features on preference ratings across different landscape types. In Kardan and Berman et al. (2015)’s work, they suggested that lower hue, higher edge, and more variations in saturation were related to more preference. Similar results were not found in our study. This discrepancy between our results and previous literature could be caused by the construction of our videos. As shown in the correlation matrix (see above), the saturation and SD saturation in our generated videos have a very high correlation with each other, and the brightness and SD brightness also have a very high correlation. In general, both saturation and SD saturation, and brightness and SD brightness are associated positively with each other but should not be identical with each other. The high correlations indicated a potential problem of the videos we created that they might not be good representatives of scenes that we would encounter in daily life. Therefore, the predictive directions that low level visual features suggested in this study might not be applicable to real life.

Our results showed that the predictive values of semantic features of two features (billboard and lake) were congruent with what suggested in the previous literature (e.g., Nelson, 1997; Antonson et al., 2009). Artificial feature billboard negatively predicted the preference whereas the natural feature lake positively predicted the preference. However, the predictive value of trees was varied in different landscape types. This result did not necessarily suggest that people did not like trees in the environment, but instead it might suggest specific properties of trees that people would prefer. There might be three possible properties that could affect the influences of trees on the preference: density, openness, and variations in the environment. However, trees were only measured by objective prevalence (an algorithm calculated the relative tree-greenness) in this study and we did not have a validate measure for either density or openness. Therefore, we could not further confirm our speculations of the preference of trees in the environment.

To conclude, our results suggest that low-level visual features and semantic features can reliably predict aesthetic preferences, in which lake has a positive impact on scenic beauty whereas billboard has a negative impact. The predictive values of those features might vary depending on the types of landscape and environmental elements presented. Further analysis should be conducted with the videos generated in the aim of only manipulating one low-level visual feature at each time in order to have a more nuanced and precise understanding of how each of low-level visual features affect aesthetic preferences. Future research should also include more types as well as measurements of semantic features in order to be able to quantify and explain what aspects of the semantic features can influence aesthetic preferences and how they interact with low-level visual features in predicting preference ratings. By having a more concrete and nuanced result of the predictive values of low-level visual features and semantic features on preference, we can better advise urban environment designers and architects on ways to construct an aesthetically pleasing environment.